CN109947898B

CN109947898B - Equipment fault testing method based on intellectualization

Info

Publication number: CN109947898B
Application number: CN201811336032.3A
Authority: CN
Inventors: 方冰; 张翠侠
Original assignee: CETC 28 Research Institute
Current assignee: CETC 28 Research Institute
Priority date: 2018-11-09
Filing date: 2018-11-09
Publication date: 2021-03-05
Anticipated expiration: 2038-11-09
Also published as: CN109947898A

Abstract

The invention relates to an equipment fault prediction method based on intellectualization, which utilizes the original data and mastered knowledge of the equipment collected at present and comprehensively utilizes two intellectualized fault prediction methods based on rule driving and data driving to quickly predict whether the equipment has faults or not and locate the faults from a large amount of guarantee data, thereby improving the timeliness and the accuracy of fault prediction and quickly and efficiently assisting the guarantee personnel to accurately guarantee an equipment system.

Description

Equipment fault testing method based on intellectualization

Technical Field

The invention relates to an equipment fault testing method based on intellectualization.

Background

At present, the fault detection method of the existing equipment can not detect potential faults, and the faults can be detected only when the faults develop to a certain stage, so that 'after repair' is carried out. When the equipment is used, the equipment is maintained regularly, and maintenance is carried out after the fault is found, namely, maintenance modes of 'planned maintenance', 'after-the-fact maintenance' and 'planned maintenance', so that catastrophic faults are difficult to prevent. The equipment failure prediction method based on intellectualization excavates the intrinsic knowledge and rules of equipment guarantee data through two methods based on rule driving and based on data driving, and predicts when and what kind of failure occurs in the equipment so as to take timely and effective prevention and maintenance measures and ensure the normal execution of tasks. The intelligent fault prediction method provides a basis for improving the timeliness and the accuracy of fault prediction.

Disclosure of Invention

The purpose of the invention is as follows: in view of the above problems, it is an object of the present invention to provide a method for testing equipment failure based on that only.

The technical scheme is as follows: in order to achieve the above purpose, the equipment fault testing method based on intellectualization provided by the invention comprises the following steps:

(1) a rule-driven fault prediction method;

(2) a data-driven based fault prediction method;

(3) under the condition that an object system has a large amount of prior knowledge and a relatively complete knowledge rule base, a fault prediction method based on rule driving is adopted; and in the case that the expert experience is lacked, an accurate mathematical analysis model cannot be established, and only relevant data such as equipment attribute, performance, dynamics and the like can be utilized when various types of faults occur, a fault prediction party based on data driving is preferentially selected.

The two fault prediction methods can be comprehensively utilized according to actual conditions, and the advantages of the two fault prediction methods are combined to continuously learn from experience and feed back circularly, so that the timeliness and the accuracy of fault prediction are improved.

In the step 1, the specific content of the fault prediction method based on rule driving is as follows:

(1) establishing a knowledge graph: modeling the relation between knowledge and knowledge by adopting a graph, wherein nodes in the graph represent entities of the knowledge, edges on the graph represent the relation between the entities, and entity labels are adopted to distinguish the entities; the knowledge graph construction mainly comprises knowledge extraction, knowledge fusion and knowledge reasoning; the knowledge extraction is to extract and identify equipment knowledge data from multi-source heterogeneous data, wherein the equipment knowledge data comprises equipment entities, corresponding relations and attribute knowledge elements; knowledge fusion, namely multi-source heterogeneous data fusion, wherein the knowledge of multi-source heterogeneity, semantic diversity and dynamic evolution acquired from fragmented data is subjected to correctness judgment by adopting a conflict detection and consistency check method, and useful information is picked to organize into a knowledge base;

(2) inducing inference rules by adopting first-order predicate logic; the rules are expressed by ECA model, each rule in the rule base is composed of trigger time, rule conditions and rule actions, and when a trigger event occurs, corresponding action is executed according to the condition of meeting the condition conditions of the rules.

And G denotes a knowledge graph, G ═ (E, R, F), where E is the set of entities in the knowledge base, R is the set of relationships in the knowledge base,

represents facts in the knowledge base, each fact may be composed of a triple, entity 1, relationship, entity 2; the inference rule is induced by adopting first-order predicate logic,

in (1), CaptialOf, LocatedIn represents predicates, x, y represent individual variables, and logic implications

Meaning "if …, then …", a full-scale word

Meaning "to arbitrary," "to whatever," or "to all" semantics; LocatedIn (x, y) is expressed as a rule header, which is the premise of the rule, CaptialOf (x, y) is expressed as a rule body, which is the conclusion of the rule, and matching and reasoning are realized between the two through a rule engine.

The entity tag is tag information having an identifying meaning in an entity.

The knowledge graph is formed by extracting knowledge reasoning through a rule formulated in advance, and a rule or axiom can be additionally added to express a more complex constraint relation.

The correct knowledge is organically organized into a knowledge base through methods such as entity alignment, attribute alignment, normalization and the like, and comprehensive knowledge sharing is provided.

In the step 2, the specific content of the fault prediction method based on data driving is as follows:

(1) constructing a depth model which comprises a plurality of nerve layers, wherein each nerve layer consists of a plurality of sensors, and each sensor comprises a group of input nodes, an output node, an activation function and a group of weights; let the input vector be x, the output value be y: the output result of the sensor is:

y＝activation(w*x+b) (1)

wherein w is a weight vector, b is a bias vector, and activation is an activation function;

the activation function and the weight vector map the input vector to a single scalar output value; the value range of the output value y is determined according to the activation function, and the output of the sensor approaches to the output value through training and learning of the weight vector w and the bias vector b;

(2) performing a feedforward process by an error back propagation algorithm to obtain a model output, then calculating a cost function, and correcting a weight matrix layer by layer from an output layer by using a gradient descent method, wherein the weight calculation of each layer of neuron must depend on the weight calculation result of the previous layer; since the error flow in this process is from top to bottom, this training algorithm is called an error back propagation algorithm.

The number of layers of the deep neural network is three, and the input parameter is x₁、x₂And x₃The output value of the second layer is

Output node is H_w,b(x) (ii) a Predicting a neural network graph based on the data-driven faults;

for the output of the second layer

Comprises the following steps:

wherein

Linear coefficients of the 1 st neuron, the 2 nd neuron and the 3 rd neuron of the second layer to the 1 st neuron of the third layer, respectively; and so on,

linear coefficients of the 1 st neuron, the 2 nd neuron and the 3 rd neuron of the second layer to the 2 nd neuron of the third layer, respectively,

linear coefficients of the 1 st neuron, the 2 nd neuron and the 3 rd neuron of the second layer to the 3 rd neuron of the third layer, respectively;

represents the bias amounts of the 1 st neuron, the 2 nd neuron and the 3 rd neuron of the second layer, respectively;

for the output node, there are:

assuming that there are m neurons in layer l-1, then for the output of the jth neuron in layer l, there are:

can also be expressed by a matrix method as

a^l＝σ(z^l)＝σ(W^la^l-1+b^l) (5)。

Step 3 comparing the two failure prediction methods

According to the description of two prediction methods of the fault prediction based on rule driving and the fault prediction based on data driving, through analysis and comparison, the following can be seen: the two failure prediction methods have advantages and disadvantages respectively, and the comparison result is shown in table 1.

TABLE 1 comparison of two failure prediction methods

The fault prediction scheme based on rule driving has strong interpretability, but a complete knowledge rule base is difficult to construct; the fault prediction method based on data driving does not need prior knowledge of an object system, can perform fault prediction by analyzing and mining implicit information in data on the basis of the data, but depends on the data, and has weak interpretability. The two fault prediction methods are comprehensively utilized, and the advantages of the two methods are combined, so that the continuous learning and the cyclic feedback are realized from experience, and the timeliness and the accuracy of the fault prediction are improved.

Has the advantages that: compared with the prior art, the invention has the following advantages:

1. the method has the advantages that the method utilizes the original data and mastery knowledge of the equipment collected at present, comprehensively utilizes two intelligent fault prediction methods based on rule driving and data driving, quickly predicts whether the equipment has faults or not and positions the faults from a large amount of guarantee data, improves timeliness and accuracy of fault prediction, and quickly and efficiently assists the guarantee personnel in accurately guaranteeing the equipment system.

Drawings

FIG. 1 is an equipment age knowledgegraph;

FIG. 2 is an equipment failure knowledge map;

FIG. 3 is an equipment system knowledge graph;

FIG. 4 is a diagram of a model of a perceptron;

FIG. 5 is a diagram of a conventional neural network;

FIG. 6 is a diagram of a data-driven failure prediction neural network;

fig. 7 is a diagram of a data-driven based failure prediction process.

Detailed Description

According to the preferred embodiment of the present invention, an intelligent equipment failure prediction method includes the following 2 steps, which are shown in fig. 1 to 7 and described in detail below.

Step 1, fault prediction method based on rule driving

In the step, taking an equipment system as an example, according to an equipment system model, carrying out knowledge fusion and knowledge reasoning on structured and unstructured information, constructing an equipment system knowledge graph and predicting equipment fault information based on the graph, wherein the equipment system knowledge graph is shown in FIG. 3;

step 1-1, extracting equipment entities such as missile weapons, transport vehicles and the like, attribute information such as equipment models, manufacturers, production time, working parameters and the like, and all knowledge of relationships such as deployment, guarantee and the like from multi-source equipment guarantee data through word segmentation, part-of-speech tagging, grammar and emotional tendency analysis based on basic models such as classification, clustering and graph algorithms. And modeling the equipment body based on the body construction scheme. The equipment ontology conceptual subclass can be divided into transportation vehicles, weaponry, special purpose equipment, general purpose equipment, etc., and the relationships include deployment, carriage, dynamics, etc., as shown in the upper part (mode level) of FIG. 3;

step 1-2, integrating, disambiguating, processing and verifying multi-source heterogeneous data according to the established equipment body model, and constructing an equipment knowledge graph through knowledge fusion, as shown in the lower half (data layer) of fig. 3;

step 1-3, the equipment system knowledge graph can comprise the following two types of rules: (1) aiming at the rule of the attribute, the attribute value can be obtained through numerical calculation; (2) and discovering implicit relations among the entities through chained rules aiming at the rules of the relations. According to the rule of the attribute, the knowledge graph spectrum comprises the production time and the planned scrapping time of a certain device, the current distance scrapping time or the time of the current distance scrapping can be obtained through reasoning, and the rule can be expressed as follows:

RULE < equipment remaining life > [ < equipment identification >, < equipment type >, < equipment production time >, < equipment planned scrap time >, < current time > ]

WHEN (Current time)

IF < equipment expected scrap time > current time > THEN < equipment remaining life ═ equipment expected scrap time-current time >

END-RULE [ < equipment residual life > ]

According to the above rule, it can be known

The semantic information expressed by the rule is added into the original equipment age-based knowledge graph, namely a new fact (the age of the transport vehicle A, owned equipment and scrapped equipment) is added on the basis of the figure 1(a), and the obtained new equipment age-based knowledge graph is shown in the figure 1 (b).

Deducing that the transport vehicle A is still scrapped for 2 years by combining scrapping time and current time;

rules for relationships, such as signs that one of the equipment of a same model, batch of equipment has failed and diagnosed causes of the failure, can infer that another equipment of the same phenomenon may have failed and cause of the failure;

RULE < equipment failure cause > [ < equipment identification >, < equipment type >, < equipment history failure phenomenon >, < equipment history failure cause >, < equipment current state > ]

WHEN (Current State (Equipment identification B))

IF < current state (equipment identity B) LIKE historical fault phenomenon (equipment identity A) > THEN < equipment fault reason (equipment identity B) LIKE equipment historical fault reason (equipment identity A) >

END-RULE [ < cause of equipment failure > ]

According to the above rule, it can be known

The semantic information expressed by the rule is added into the original equipment fault knowledge graph, namely a new fact (transport vehicle B, prediction, fault reason) is added on the basis of the graph in the figure 1(a), and the obtained new equipment fault knowledge graph is shown in the figure 1 (B).

And according to the rule, deducing and predicting that the components of the transport vehicle B will have faults by combining the current state of the transport vehicle B according to the historical fault phenomenon and the fault reason of the transport vehicle B.

Step 2, fault prediction method based on data driving

Taking an equipment system as an example, a method adopting deep learning is described, and a process for predicting equipment failure based on data drive is shown in fig. 7.

And 2-1, collecting data, finding out equipment fault occurrence influence factors, mainly comprising equipment self factors such as equipment attributes (such as equipment identification, equipment type, length, width and height), performance (such as maximum equipment maneuvering speed, minimum turning radius and equipment service life), deployment (such as longitude, latitude and elevation), dynamic (such as equipment maneuvering speed, equipment maneuvering direction and equipment state) and environmental factors such as weather types, temperature, relative humidity, wind degree and horizontal visibility), terrain (such as road conditions and high altitude), electromagnetic environment (such as electromagnetic frequency and power) and collecting related historical accumulated data. Assuming a random selection of data from the collected historical data, a 3-dimensional data set is created with data parameters as shown in table 2. Grouping data, dividing the collected data into 2 groups which are respectively { (v)_i,t_i,f_i),Y_iAnd { (v)_j,t_j,f_j),Y_jIn which Y is_i，Y_jAnd respectively, the label corresponding to each group of data, namely normal or fault. Multiple collection to find Y>3700 the equipment is normal, otherwise the equipment is in failure;

TABLE 2 three-dimensional data

Name (R)	Symbol
		Equipment maneuvering speed/(km/h)	v
Equipment service time/h	t
		Electromagnetic frequency/Hz	f

And 2-2, carrying out deep learning on historical data by adopting a Dynamic Bayesian Network (DBNs) model to realize layer-by-layer feature extraction, carrying out layer-by-layer training, and obtaining a training conclusion that the tire of the transport vehicle A will break down in a period of time in the future. Through superposition of a plurality of Restricted Boltzmann Machine (RBM) models, each RBM is a two-layer model only containing one hidden layer, and the training output of each RBM is used as the input of the next RBM. A plurality of RBM models form a probability generation model DBNs through bottom-up combination. The training process of the DBNs model can be regarded as a layer-by-layer feature extraction process, and the essential rule of equipment faults is abstracted through unsupervised feature training and supervised parameter fine tuning so as to serve as a new training and testing sample;

assuming that the failure prediction neural network layer based on data driving is three layers, and the input parameters are the training samples of the 3-dimensional data set selected in the step a), and the parameter x₁、x₂、x₃Respectively the maneuvering speed of the equipment, the service time of the equipment and the electromagnetic frequency, and the output value of the second layer is

Weight matrix of the second layer

Bias vector of second layer

Weight matrix W of the third layer³＝[0.5 0.3 0.2]Bias vector b of the third layer³Let us assume the equipment maneuvering speed x is 0.3₁80km/h, equipment service time x₂8760h, electromagnetic frequency x₃To 100hz, the output value of the second layer is calculated using equation (2)

Formula (3)) Calculating the third layer, i.e. the output node value

To 3818.83, the above output value can also be calculated by equations (4) and (5). The output node value is the equipment state in the equipment fault prediction based on the data driving, and if the output node value is more than 3700, the equipment can be deduced to be in the normal state;

step 2-3, comparing the training conclusion with the historical result, if the training conclusion is not consistent with the historical result, continuously correcting the training model, mainly correcting the weight vector w and the bias vector b, and retraining until the two conclusions are consistent;

step 2-4, collecting current equipment attribute, equipment performance, equipment deployment, equipment dynamic information, climate, terrain, electromagnetism and other environmental factors as input of deep learning;

2-5, predicting information such as fault generation probability and fault generation position by using the corrected training model through deep learning; if the predicted conclusion is inconsistent with the actual result, the training model still needs to be modified.

Claims

1. The equipment fault testing method based on the intellectualization is characterized in that: the method comprises the following steps:

(1) a rule-driven fault prediction method;

(2) a data-driven based fault prediction method;

(3) under the condition that an object system has a large amount of prior knowledge and a relatively complete knowledge rule base, a fault prediction method based on rule driving is adopted; the method is characterized in that a data-driven fault prediction method is preferentially selected under the conditions that expert experience is lacked, an accurate mathematical analysis model cannot be established at the same time, and only equipment attribute, performance and dynamic related data are available when various types of faults occur;

(2) inducing inference rules by adopting first-order predicate logic; the rules are expressed by ECA model, each rule in the rule base is composed of trigger time, rule conditions and rule actions, and when a trigger event occurs, corresponding action is executed according to the condition of meeting the condition conditions of the rules;

y＝activation(w*x+b) (1)

2. The intelligent-based equipment fault testing method of claim 1, wherein: and G denotes a knowledge graph, G ═ (E, R, F), where E is the set of entities in the knowledge base, R is the set of relationships in the knowledge base,

Meaning "if …, then …", a full-scale word

3. The intelligent-based equipment fault testing method of claim 2, wherein: the number of the nerve layers is three, and the input parameter is x₁、x₂And x₃The output value of the second layer is

Output node is H_w,b(x) (ii) a Predicting a neural network based on the data-driven failure;

for the output of the second layer

Comprises the following steps:

wherein

for the output node, there are:

can also be expressed by a matrix method as

a^l＝σ(z^l)＝σ(W^la^l-1+b^l) (5)。